Our solution approached the problem as a text classification task rather than a structured JSON generation problem.
\nInstead of directly predicting the complete expected JSON output, we observed that the first item in the target JSON often represented the primary product requested in the customer order. We therefore extracted this product name and trained a classifier to predict it directly from the natural language order text.
\nThe final pipeline consisted of:
\nThis approach was intentionally simple, computationally efficient, and easy to reproduce.
\nThe training dataset contained:
\nEach training example included:
\norder (customer request in natural language)expected_json (structured target output)We extracted the first item name from the JSON structure and used it as the prediction target.
\nExample:
\n{\n \"items\": [\n {\n \"name\": \"Caramel Macchiato\"\n }\n ]\n}\nbecame:
\nLabel = \"Caramel Macchiato\"\nThe extraction was implemented using a simple JSON parser.
\nAfter extraction, the dataset contained multiple beverage and food categories such as:
\nThe label distribution was relatively balanced, making multiclass classification feasible without special re-weighting.
\nOrder text was transformed using TF-IDF features.
\nConfiguration:
\nTfidfVectorizer(\n ngram_range=(1, 2),\n max_features=30000,\n stop_words=\"english\"\n)\nKey choices:
\nCapture individual menu keywords:
\nlatte\namericano\nbagel\nmocha\nCapture menu phrases:
\nflat white\nchai latte\ncaramel macchiato\nBigrams were particularly important because many products are naturally expressed as multi-word names.
\nWe trained a multiclass Logistic Regression model:
\nLogisticRegression(\n max_iter=1000,\n n_jobs=-1\n)\nReasons for choosing Logistic Regression:
\nThe model learns associations between words in the order text and product labels.
\nFor example:
\n\"large caramel macchiato\"\nis mapped toward the
\nCaramel Macchiato\nclass.
\nThe final prediction pipeline was:
\nOrder Text\n ↓\nTF-IDF Vectorization\n ↓\nLogistic Regression\n ↓\nPredicted Product\n ↓\nSubmission File\nThe trained model was applied to all test orders, and predicted product names were written into the submission file.
\nSeveral properties of the dataset favored this approach:
\nProduct names were often explicitly mentioned in customer orders.
TF-IDF effectively captured discriminative menu keywords.
Logistic Regression performs well on sparse text representations.
The training set was relatively small, making lightweight linear models competitive.
This solution has several limitations:
\nFuture work could involve:
\nWe transformed the problem into a multiclass text classification task by extracting the primary item from each target JSON. Using TF-IDF features and Logistic Regression, we obtained a simple, efficient, and interpretable solution that served as a strong baseline for the competition.
\nOur solution approached the problem as a text classification task rather than a structured JSON generation problem.
\nInstead of directly predicting the complete expected JSON output, we observed that the first item in the target JSON often represented the primary product requested in the customer order. We therefore extracted this product name and trained a classifier to predict it directly from the natural language order text.
\nThe final pipeline consisted of:
\nThis approach was intentionally simple, computationally efficient, and easy to reproduce.
\nThe training dataset contained:
\nEach training example included:
\norder (customer request in natural language)expected_json (structured target output)We extracted the first item name from the JSON structure and used it as the prediction target.
\nExample:
\n{\n \"items\": [\n {\n \"name\": \"Caramel Macchiato\"\n }\n ]\n}\nbecame:
\nLabel = \"Caramel Macchiato\"\nThe extraction was implemented using a simple JSON parser.
\nAfter extraction, the dataset contained multiple beverage and food categories such as:
\nThe label distribution was relatively balanced, making multiclass classification feasible without special re-weighting.
\nOrder text was transformed using TF-IDF features.
\nConfiguration:
\nTfidfVectorizer(\n ngram_range=(1, 2),\n max_features=30000,\n stop_words=\"english\"\n)\nKey choices:
\nCapture individual menu keywords:
\nlatte\namericano\nbagel\nmocha\nCapture menu phrases:
\nflat white\nchai latte\ncaramel macchiato\nBigrams were particularly important because many products are naturally expressed as multi-word names.
\nWe trained a multiclass Logistic Regression model:
\nLogisticRegression(\n max_iter=1000,\n n_jobs=-1\n)\nReasons for choosing Logistic Regression:
\nThe model learns associations between words in the order text and product labels.
\nFor example:
\n\"large caramel macchiato\"\nis mapped toward the
\nCaramel Macchiato\nclass.
\nThe final prediction pipeline was:
\nOrder Text\n ↓\nTF-IDF Vectorization\n ↓\nLogistic Regression\n ↓\nPredicted Product\n ↓\nSubmission File\nThe trained model was applied to all test orders, and predicted product names were written into the submission file.
\nSeveral properties of the dataset favored this approach:
\nProduct names were often explicitly mentioned in customer orders.
TF-IDF effectively captured discriminative menu keywords.
Logistic Regression performs well on sparse text representations.
The training set was relatively small, making lightweight linear models competitive.
This solution has several limitations:
\nFuture work could involve:
\nWe transformed the problem into a multiclass text classification task by extracting the primary item from each target JSON. Using TF-IDF features and Logistic Regression, we obtained a simple, efficient, and interpretable solution that served as a strong baseline for the competition.
\n